{
  "id": 8161,
  "title": "Normalization in Vowpal-Wabbit",
  "url": "/competitions/acquire-valued-shoppers-challenge/discussion/8161",
  "author_name": "",
  "post_date": "2014-05-15T06:38:38.543Z",
  "votes": 1,
  "comment_count": 1,
  "views": 1916,
  "content": "<p>Vowpal-Wabbit will normalize the data by default. Does it means that we do not have to normalize it by ourself?</p>\n\n<p>Any one knows how VW normalizes data ?</p>",
  "messages": [
    {
      "id": "44578",
      "postDate": "05/15/2014 06:38:38",
      "content": "<p>Vowpal-Wabbit will normalize the data by default. Does it means that we do not have to normalize it by ourself?</p>\n\n<p>Any one knows how VW normalizes data ?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "44583",
      "postDate": "05/15/2014 07:57:51",
      "content": "<p>you can read it up here:&nbsp;https://www.cs.cmu.edu/~sross1/publications/uai13_normalized.pdf</p>\n<p>As far as I know, it basically maintains&nbsp;scale parameters for all features to&nbsp;record how large the features' values&nbsp;can get. At each step in gradient descent, it checks if&nbsp;the scale parameters need updates. If yes, the corresponding weights are updated to the values they would have assumed had that new scale value been used in the first place.</p>\n<p>In the paper, the authors also did some experiments in which this technique&nbsp;delivered results comparable to those you would get by putting normalized data into vw.</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 44583,
      "author_name": "breuker",
      "author_url": "",
      "post_date": "05/15/2014 07:57:51",
      "content": "<p>you can read it up here:&nbsp;https://www.cs.cmu.edu/~sross1/publications/uai13_normalized.pdf</p>\n<p>As far as I know, it basically maintains&nbsp;scale parameters for all features to&nbsp;record how large the features' values&nbsp;can get. At each step in gradient descent, it checks if&nbsp;the scale parameters need updates. If yes, the corresponding weights are updated to the values they would have assumed had that new scale value been used in the first place.</p>\n<p>In the paper, the authors also did some experiments in which this technique&nbsp;delivered results comparable to those you would get by putting normalized data into vw.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "44578": "",
    "44583": ""
  },
  "source": "meta"
}